Message-ID: <672228a5-5b5b-4c32-9fd9-4e1b59c08dd2@arm.com>
Date: Thu, 14 Mar 2024 11:32:34 +0100
From: Pierre Gondois <pierre.gondois@....com>
To: Ze Gao <zegao2021@...il.com>
Cc: linux-kernel@...r.kernel.org, Peter Zijlstra <peterz@...radead.org>,
 Mel Gorman <mgorman@...e.de>, Juri Lelli <juri.lelli@...hat.com>,
 Ingo Molnar <mingo@...hat.com>, Dietmar Eggemann <dietmar.eggemann@....com>,
 Daniel Bristot de Oliveira <bristot@...hat.com>,
 Ben Segall <bsegall@...gle.com>, Vincent Guittot
 <vincent.guittot@...aro.org>, Valentin Schneider <vschneid@...hat.com>,
 Steven Rostedt <rostedt@...dmis.org>, Ze Gao <zegao@...cent.com>
Subject: Re: [PATCH] sched: Improve rq selection for a blocked task when its
 affinity changes



On 3/14/24 04:49, Ze Gao wrote:
> On Thu, Mar 14, 2024 at 10:59 AM Ze Gao <zegao2021@...il.com> wrote:
>>
>> On Wed, Mar 13, 2024 at 11:16 PM Pierre Gondois <pierre.gondois@....com> wrote:
>>>
>>> Hello Ze,
>>>
>>> I am running stress-ng with the following command:
>>>     stress-ng -c 1 -l 10 &
>>> and migrating the process with:
>>>     taskset -pc [cpus] [pid]
>>>
>>> The thread seems to be migrated via:
>>> sched_setaffinity
>>>     \-__sched_setaffinity()
>>>       \-__set_cpus_allowed_ptr()
>>>         \-__set_cpus_allowed_ptr_locked()
>>>           \- [1]
>>
>>
>>
>>> [1]
>>> /*
>>>    * Picking a ~random cpu helps in cases where we are changing affinity
>>>    * for groups of tasks (ie. cpuset), so that load balancing is not
>>>    * immediately required to distribute the tasks within their new mask.
>>>    */
>>> dest_cpu = cpumask_any_and_distribute(cpu_valid_mask, ctx->new_mask);
>>>
>>> So it seems the destination CPU chosen among the new CPU affinity mask is done
>>> here, by picking a random CPU in the mask.
>>
>> IIUC, this is for running/queued/waking tasks instead of blocked tasks.
>>
>> Am I missing something obvious here?

Not at all, I just didn't tested the patch on a blocked task as you said.

>>
>>> Checking the cpus_ptr in select_idle_sibling() might be useful in other cases,
>>> but I think the experiment doesn't show that. Maybe a another small tweak could
>>
>> The experiment is used to illustrate that the status quo does not do well
>> but has to rely on select_fallback_rq() to choose a cpu for a woken task
>> which turns out to be a bad choice since it's already monopolized by a
>> cpu bound task, that is why a second migration happens with the help
>> of the load balancer.
> 
> Btw, perf alone does not show obvious results here, you need some
> other observability tools to make sure the migration is not initiated by
> __set_cpus_allowed_ptr_locked (i.e., for running tasks). I achieve this
> by directly adding some tracepoints to both select_fallback_rq() and
> select_idle_sibling().

FWIW, I could effectively reproduce the issue described in the commit message, so:
Tested-by: Pierre Gondois <pierre.gondois@....com>

Regards,
Pierre

> 
> Regards,
>          -- Ze
> 
>> Actually, we can reuse the same reasons for doing so as in
>>
>>      commit 46a87b3851f0("sched/core: Distribute tasks within affinity masks")
>>
>>> be done at [1] instead ?
>>
>> As for blocked tasks, check out what is commented on set_task_cpu() and
>> select_task_rq(), since we never call set_task_cpu() on blocked tasks which
>> in turn, we have no way to change p->wake_cpu to dest_cpu being randomly
>> chosen here, so when it's woken up, it still needs to go through the
>> select_task_rq() process using the outdated p->wake_cpu.
>>
>>
>> Thanks,
>>          -- Ze
>>
>>> Regards,
>>> Pierre
>>>
>>> On 3/13/24 09:58, Ze Gao wrote:
>>>> We observered select_idle_sibling() is likely to return the *target* cpu
>>>> early which is likely to be the previous cpu this task is running on even
>>>> when it's actually not within the affinity list newly set, from where after
>>>> we can only rely on select_fallback_rq() to choose one for us at its will
>>>> (the first valid mostly for now).
>>>>
>>>> However, the one chosen by select_fallback_rq() is highly likely not a
>>>> good enough candidate, sometimes it has to rely on load balancer to kick
>>>> in to place itself to a better cpu, which adds one or more unnecessary
>>>> migrations in no doubt. For example, this is what I get when I move task
>>>> 3964 to cpu 23-24 where cpu 23 has a cpu bound work pinned already:
>>>>
>>>>           swapper       0 [013]   959.791829: sched:sched_migrate_task: comm=stress-ng-cpu pid=3964 prio=120 orig_cpu=13 dest_cpu=23
>>>> kworker/24:2-mm    1014 [024]   959.806148: sched:sched_migrate_task: comm=stress-ng-cpu pid=3964 prio=120 orig_cpu=23 dest_cpu=24
>>>>
>>>> The thing is we can actually do better if we do checks early and take more
>>>> advantages of the *target* in select_idle_sibling(). That is, we continue
>>>> the idle cpu selection if *target* fails the test of cpumask_test_cpu(
>>>> *target*, p->cpus_ptr). By doing so, we are likely to pick a good candidate,
>>>> especially when the newly allowed cpu set shares some cpu resources with
>>>> *target*.
>>>>
>>>> And with this change, we clearly see the improvement when I move task 3964
>>>> to cpu 25-26 where cpu 25 has a cpu bound work pinned already.
>>>>
>>>>           swapper       0 [027]  4249.204658: sched:sched_migrate_task: comm=stress-ng-cpu pid=3964 prio=120 orig_cpu=27 dest_cpu=26
>>>>
>>>> Note we do the same check for *prev* in select_idle_sibling() as well.
>>>>
>>>> Signed-off-by: Ze Gao <zegao@...cent.com>
>>>> ---
>>>>    kernel/sched/fair.c | 13 ++++++++-----
>>>>    1 file changed, 8 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
>>>> index 533547e3c90a..9ef6e74c6b2a 100644
>>>> --- a/kernel/sched/fair.c
>>>> +++ b/kernel/sched/fair.c
>>>> @@ -7511,16 +7511,19 @@ static int select_idle_sibling(struct task_struct *p, int prev, int target)
>>>>         */
>>>>        lockdep_assert_irqs_disabled();
>>>>
>>>> -     if ((available_idle_cpu(target) || sched_idle_cpu(target)) &&
>>>> -         asym_fits_cpu(task_util, util_min, util_max, target))
>>>> +     if (cpumask_test_cpu(target, p->cpus_ptr) &&
>>>> +             (available_idle_cpu(target) || sched_idle_cpu(target)) &&
>>>> +             asym_fits_cpu(task_util, util_min, util_max, target))
>>>>                return target;
>>>>
>>>>        /*
>>>>         * If the previous CPU is cache affine and idle, don't be stupid:
>>>>         */
>>>> -     if (prev != target && cpus_share_cache(prev, target) &&
>>>> -         (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
>>>> -         asym_fits_cpu(task_util, util_min, util_max, prev)) {
>>>> +     if (cpumask_test_cpu(prev, p->cpus_ptr) &&
>>>> +             prev != target &&
>>>> +             cpus_share_cache(prev, target) &&
>>>> +             (available_idle_cpu(prev) || sched_idle_cpu(prev)) &&
>>>> +             asym_fits_cpu(task_util, util_min, util_max, prev)) {
>>>>
>>>>                if (!static_branch_unlikely(&sched_cluster_active) ||
>>>>                    cpus_share_resources(prev, target))

Powered by blists - more mailing lists